Lidar-based individual tree classification using convolutional neural network

• Main Authors: Hsuan-Tsung Chang, 張鉉宗
• Other Authors: Chao-Cheng Wu
• Format: Others
• Language: en_US
• Published: 2018
• Online Access: http://ndltd.ncl.edu.tw/handle/9r66u3

碩士 === 國立臺北科技大學 === 電機工程系 === 107 === Tree conservation and forest management play a significant role in the ecosystem. Remote sensing techniques provide an efficient way to collect information about forest stands on a massive scale. Especially the terrain features in mountainous areas make the forest survey become very difficult and time-consuming. Light Detection and Ranging (LiDAR) data have emerged as new resource for detecting forest stands and estimating tree height. High-resolution LiDAR image contributes to the development of many algorithms for extracting individual tree information. Initial parameters or thresholds are one of the major issues of individual tree detection. Some algorithms determined these parameters based on image processing method, such as local maximum filtering (Wulder et al., 2000; Pouliot et al., 2002), template matching (Pollock, 1996), etc. Nevertheless, these algorithms still need prior or posterior knowledge to enhance accuracy or reduce error of commission through experiments. Besides, parameters or thresholds would be affected by growth conditions of natural forest and sampling rate of sensors. This thesis utilized a convolutional neural network (CNN) classifier for tree detection algorithm on LiDAR image. Regions of interest were determined by the modified local maximum algorithm in the first stage. Then, the features would be learned by the CNN automatically for classification of trees. The experimental results demonstrated that the features extracted by CNN provide good classification result to recognize the non-tree areas and forest stands